Peng Qian (@qtalen) header
Peng Qian (@qtalen) profile picture

I was with Alibaba as a senior data scientist , and now I am the chief architect in the big data field at a major brokerage firm in China.

Joined April 2023
Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)Photo by Peng Qian (@qtalen)
I spent over a month reworking the entire team's AI Coding workflow. We brought in Amazon's aidlc-workflows and adapted it to fit our own needs. Both the team's AI development efficiency and code quality improved significantly.
13
12
1,530
Peng Qian retweeted
今天 GitHub 日榜:AI 审美觉醒日,设计 Skill 集体起义,专治"一眼 AI 味"! 5 个星标暴增最狠的项目,专业拆解下! 1. pbakaus/impeccable 让 AI 编程助手更懂设计的设计语言。 1 个 skill + 24 条命令 + 浏览器实时迭代 + 62 条确定性检测规则,专抓紫蓝渐变、卡片套卡片这些 AI 前端通病。 🔗 直达:github.com/pbakaus/impeccabl… AI 写的页面终于不再千篇一律"Inter 字体 + 紫色渐变"了。619 星一天,总星冲破 8.6 万,审美革命正式打响! 🟢 2. makecindy/cindy 开源、开箱即用的 AI Agent。 把多种 harness、模型和工具整合进一个 Agent,在你自己的电脑上真刀真枪干活,桌面端、移动端全覆盖。 🔗 直达:github.com/makecindy/cindy "想到,就能做到",这口号够狂,Agent 已经从玩具变成打工人。477 星一天,个人 Agent 赛道又来猛将! 🟢🟢 3. Nutlope/hallmark 拒绝 AI 味的设计 Skill,支持 Claude Code、Codex、Cursor、OpenCode。 48 套主题随机发牌,在真实浏览器里逐条检查"机器味"痕迹,最后还有全新上下文的评审把关,Together AI 出品。 🔗 直达:github.com/Nutlope/hallmark 和 impeccable 同天上榜,反 AI-slop 统一战线成立了!455 星一天,设计师们可以松口气了。 🟢🟢🟢 4. luongnv89/claude-howto 可视化、案例驱动的 Claude Code 实战指南。 从基础概念一路讲到高级 Agent,附可直接复制的模板,还有中文版。 🔗 直达:github.com/luongnv89/claude-… "一个周末精通 Claude Code",这是给新手的速成秘籍。392 星一天,总星 4.4 万,学习党狂喜! 🟢🟢🟢🟢 5. pollen-robotics/microduck 一只会走路的迷你双足鸭子机器人 🦆 这个仓库是鸭子的"大脑":约 25cm、800g,Rust 写的 50Hz 控制循环用强化学习策略驱动 15 个舵机。 🔗 直达:github.com/pollen-robotics/m… 具身智能也能这么可爱?RL 训练出来的鸭子步伐,萌到犯规。319 星一天,硬核与萌系兼得! ⚠️⚠️ 总结: 今天的信号很清楚:AI 能写代码只是起点,写出"有品味"的东西才是下一个分水岭。设计 Skill + 开箱 Agent,就是 2026 的新基建!
43
139
17
1,064
10,067,493
Peng Qian retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source! Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable. 🔗 github.com/deepseek-ai/DeepI… 1/4
1,771
4,938
2,352
43,862
7,135,606
Most AI coding workflows start to drift once the context gets too long. I introduced a graph engine based on a state machine to the AIDLC workflow. This engine guides the workflow's execution, which solves the drift problem. dataleadsfuture.com/your-ai-… #ai #vibecoding #coding
1
62
From OpenSpec to AIDLC: How I Improved My Team’s AI Code Quality by @qtalen levelup.gitconnected.com/fro…
1
1
58
Each of us has a different set of strengths, and a top-tier programmer isn't necessarily better than an average one when it comes to understanding the finer details of a business. For a dev team, the real focus should be on getting everyone into the right role, not on trying to build a team of superhumans. But most Coding Agents and SDD frameworks out there tend to push one person to do everything, from requirements analysis all the way to product ops. That just doesn't make sense. So lately I've been working on building our own process-driven development workflow on top of Amazon's aidlc workflow. The goal is to hand off different parts of the product development process to the people who actually know that part best. The docs and code are still written by AI, but people are only responsible for planning and reviewing the parts they own. So far, everything's working pretty well. The docs that Aidlc produces are incredibly detailed. Even DeepSeek Flash can build a product that fully meets the requirements just by following those docs. The consistency is great too, so you don't have to worry about code and docs drifting apart after multiple rounds of iteration. The best part is that this process-oriented workflow forces every round of human-AI interaction to leave a paper trail with approval records. So when a bug shows up in production, I know exactly whose door to knock on.
3
96
Peng Qian retweeted
🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta! 🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇 🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex! Check out the configuration details in our official API docs: api-docs.deepseek.com/quick_…
1,685
3,472
2,167
29,707
9,618,311
The essence of prompt engineering isn't about feeding the model more information. It's about helping both the model and humans put their limited attention in the right place. dataleadsfuture.com/why-well…
1
59
Even though DeepSeek-V4 doesn't have the ability to read images, I built a small OpenCode plugin that successfully gave DeepSeek indirect image reading capabilities. dataleadsfuture.com/deepseek… #ai #vibecoding #coding
59
Does your AI always write garbage code? That's because your spec files have never been properly reviewed. By adding a reflection mechanism to the proposal stage of OpenSpec, I got DeepSeek-v4-pro to write code on par with Opus 4.6: dataleadsfuture.com/reflecti… #vibecoding #ai
71
Ride the wave of AI coding, don't get swept away by it. In my latest article, I dive into the practical details of building your own personal AI coding setup using OpenCode + Oh-My-OpenCode-Slim + OpenSpec. dataleadsfuture.com/how-i-us…
1
142
The hardest part of enterprise AI agents isn’t the LLM. It’s wiring into real workflows. I built a setup where: ✅ Agent Skills load in real time from a database ✅Scripts run safely in containers ✅A “skills agent” acts as a tool to keep the main agent’s context clean
1
2
42
Most “agent memory” setups just save everything and hope semantic search will fix it. I share a pattern that works better: ✅ Use an LLM to decide what is worth remembering ✅ Store memories in RedisVL ✅ Run memory extraction in parallel with asyncio #AI #LLM #AIAgents #Redis
1
1
45
Tried RedisVL as a memory layer for AI agents: ✅ Short-term memory with MessageHistory: works great ✅ Long-term semantic memory with SemanticMessageHistory: not so much If you are thinking about RedisVL for agent memory, this will save you time: dataleadsfuture.com/build-lo…
1
32
AI agents can be smart — but without compliance guardrails, they can be risky. I used Microsoft Agent Framework middleware + AG‑UI to block unsafe prompts in a live chatbot. More on the setup + real example 👇 dataleadsfuture.com/microsof… #AI #AIAgents #LLM #DataScience
1
44